Model Choice for Spatial Prediction of Multiple Air Pollution Exposures

• Activity Number: 230
• Type: Topic Contributed
• Date/Time: Monday, July 30, 2012 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistics and the Environment
• Author(s): Adam A Szpiro*+ and Christopher J Paciorek
• Companies: University of Washington and University of California at Berkeley
• Keywords: Environmental Epidemiology ; Prediction Modeling ; Spatial Statistics ; Measurement Error ; Air Pollution ; Multi-Pollutant

Abstract

In epidemiologic studies of multiple environmental risk factors, exposures are often predicted from first-stage exposure models and plugged into second stage disease models to estimate health effect parameters of interest. For example, first-stage air pollution models use spatial statistics to predict concentrations at subject locations based on monitoring data from distinct locations. Differences between the predicted exposures and unmeasured true values result in a complex form of measurement error. Variation in the magnitudes and types of spatial prediction error between pollutants may transfer health effects from one pollutant to another (a form of measurement error induced confounding). We describe a strategy for constructing multi-pollutant exposure models that mitigates measurement error induced confounding by enforcing spatial compatibility between models for different pollutants. A notable feature of our approach is that it does not necessarily lead to exposure models with optimal predictive performance. We recommend selecting exposure models to optimize estimation of disease model parameters, rather than the intermediate step of accurately predicting the exposures.